Yuriy Poteshkin

MD, phD, endocrinologist, healthcare manager and entrepreneur

[Yuriy Poteshkin @drpoteshkin]

https://t.co/APAHpZuJxp

[Yuriy Poteshkin @drpoteshkin]

Longevity has no operating system. Eternal Search maps evidence, projects, people and capital, then uses AI to help serious teams find the funding and connections that can move life-extension research forward.
#longevity #lifeextension

[Alexey Strygin @strygah]

🚨 Anthropic is looking to acquire companies in AI x bio

I inferred this from a fresh job posting that literally describes their strategy. They want someone who can 'structure and negotiate complex deals' and 'tell a durable AI-native business from a thin wrapper.'

From the terms they use ('tuck-in acquisitions', 'acquihires'), we can tell that they're targeting rather small and medium-sized teams. But we can't be 100% sure here, because a decacorn looks kind of small next to a juggernaut aiming for a $2 trillion IPO.

Named target areas: drug discovery, clinical development, regulatory and medical writing, lab automation, healthcare data infrastructure.

[Yuriy Poteshkin @drpoteshkin]

I think I’m reasonably good at self-criticism and at seeing reality for what it is. So let me briefly clarify the point of my previous post.

The way I described it is how many people now gather information with the help of AI. Any criticism of AI on my part was there primarily to illustrate the problem. I fully understand that asking simple questions—even when the person asking them is a professional—may lead nowhere.

When patients come to me after studying their own conditions, what they usually have is a collection of fragmented facts, however purposeful and specific their questions may have been. The specialist’s job is to structure those facts and arrange them into a coherent logical framework.

But in reality, we have already moved into unexplored territory, just as I said before.

What happens when we do not know the answers, no books have been written about the subject, and the necessary experiments have never been conducted?

What happens when the question itself is one that very few people in the world are even asking?

How can we extend human life?

Or how can we stop dying altogether?

To make progress, we first need to ask ourselves what possible routes are available. In my subjective view, there are two:

Theory first, followed by experimentation.
Experiment first, followed by a theoretical explanation.

If we were dealing with conventional medicine, I would say that experimentation is clearly the way forward. It is a long and expensive route: hundreds of molecules and approaches will fail, but eventually we may find a solution.

That sounds familiar and reassuringly ordinary.

But we are not really operating within medicine, because medicine has never directly asked why humans age and die in principle. It has always approached the problem through the lens of disease.

A disease has a relatively limited clinical presentation and a set of objective signs that can be systematically identified. That framework has always been understandable. It narrows the search space, allows us to remain productive, and prevents us from going mad from the sheer number of possible explanations and interventions.

Yes, there are many diseases, just as there are probably many causes and mechanisms of ageing. Perhaps each cause will need to be controlled separately, and once we have done that, ageing itself will eventually be defeated.

But what are the actual causes—the aetiology—of ageing?

Not its hallmarks. Not its mechanisms. Its causes.

...

Silence?

All right, then what are the mechanisms of ageing?

Now suddenly there is a forest of raised hands.

So the conclusion seems obvious: if we need to intervene in the mechanisms, while the causes remain unknown and apparently difficult to identify, then we should conduct experiments.

Logical enough.

Let’s do some calculations. Before doing anything, it should at least be costed—preferably in money. :)

The OpenDrugs database contains 577 molecules with the potential to become components of an anti-ageing treatment. By studying them, we could generate new data and perhaps obtain more answers to our questions. We might even discover some of the underlying causes.

Only 577 molecules.

That does not sound like very many.

But we still have no real idea what we are dealing with. That means we would theoretically have to test every possible combination: approximately 4.9 × 10¹⁷³ possibilities.

Suppose we used an in vitro screening method, for example at IRB Barcelona. Even the automated analysis component alone would cost approximately €3.12 × 10¹⁷² to €4.89 × 10¹⁷².

And even if the screening system could process one billion combinations per second, it would still require approximately 1.57 × 10¹⁵⁷ years.

Given those numbers, I would not expect us to discover an anti-ageing drug through brute force.

And this is only 577 molecules. There will almost certainly be many more.

We need to identify promising combinations rather than test everything blindly.

At this point, one begins to wonder whether theory might be a better place to start. But experimental approaches have not yet been exhausted, so let’s continue.

Suppose we test only pairwise combinations at fixed doses.

That would give us just 166,176 combinations. I trusted AI with all these calculations. :)

We could test them across three cellular models, spend a trivial €60 million, select the single best combination, and then spend another €100 million taking it through clinical development to Phase II.

Alternatively, we could select three combinations at once and perhaps keep the total around €170 million.

This is all theoretical and based on AI’s rather casual calculations. In reality, it would probably cost more, but the relative scale is what matters.

After that, clinical development would have to proceed within a specific disease indication—because that is how the system works. “Ageing” itself is not currently accepted as a therapeutic indication—and this could cost another €500 million.

So, in principle, we might register one effective combination for around €1 billion.

But there is a complication.

We still have no idea what exactly we should be doing.

There may be a life-extending combination hidden somewhere among those 577 molecules, but nobody can guarantee that it exists.

We could therefore spend several billion dollars and obtain absolutely nothing. As we regularly see in the pharmaceutical industry, that does happen.

In other words, it may be possible to raise $1 billion from an investor—but not if we approach the matter pragmatically and admit that there are no meaningful guarantees.

At this point, I was reminded of orienteering.

On the one hand, the objective is to reach the finish line faster than everyone else. On the other hand, you must not miss a single checkpoint.

And I am exactly the kind of tedious person who refuses to miss any checkpoints.

So I suggest that we pause, take a breath, and look objectively at what is happening.

We do not have enough data to invent an effective treatment for ageing.

We also do not have adequate methods for conducting clinical trials of ageing itself.

Correct me if I am wrong. I try to remove everything I consider unnecessary, although occasionally something important gets removed with it.

To obtain data about the causes of ageing, we need to answer unanswered questions.

Questions that, apparently, nobody is even asking.

Now let us return to AI.

Can we use a combination of large language models, retrieval-augmented generation, and scripted systems to understand the current state of the field?

Yes, of course.

We can create enormous databases containing essentially all publicly available knowledge in the relevant domain.

Can such a system then critically evaluate that information and automatically debate with itself where the genuine gaps in our knowledge are?

Yes, it can.

It would require a complex multi-agent system capable of burning approximately 15 million tokens to answer a single question within the field.

In response to a query, the system could identify gaps in our knowledge and even propose experiments that might close them.

Unfortunately, humans are still much better at the experimental part.

So we would have to examine each gap manually and discuss how it could realistically be addressed.

I am writing about this as though such a system already existed.

It does.

It is called Omega Point.

All we need to do is keep the system operational. At a minimum, that requires approximately 80 million tokens per month, a programmer, servers, and other infrastructure—around $6,000 per month—so that it can continuously expand its database, answer questions, identify omissions, and propose ways to fill them.

And once our scientific council decides that a particular experiment should be conducted, we will also need funding for the experiment itself.

But how do we know that nobody has already conducted that experiment—or is conducting it right now?

How do we know that another group is not already working in exactly the same direction?

That part is simple.

For that, we have Eternal Search, which tracks what is happening across the market.

Right now, we are merging these two services into a single analytical system.

Scanning the market requires even more tokens: at least 160 million per month, in addition to the usual servers, programmers, and infrastructure. The minimum cost is approximately $8,000 per month.

We consider analytics to be the cornerstone of life-extension research. That is why we have built several systems, each designed for a different purpose but all working towards the same objective:

OpenGenes, OpenDrugs, First Approval, Omega Point, and Eternal Search.

We may get lucky and suddenly discover or understand something important.

But our plans are based on the pessimistic scenario.

And under that scenario, our objective is to find a solution using the minimum possible resources and in the shortest realistically achievable time.

In the next post, I will either explain how our analytical systems work—or why the popularisation of science may be even more important than science itself.

[Yuriy Poteshkin @drpoteshkin]

Let’s continue the discussion about AI and information.

For anyone who did not read the previous post carefully, I’ll repeat the main point: even when AI has access to all the relevant medical literature and all the available patient data, it still cannot consistently make correct clinical decisions based on that information.

That is my opinion, of course. For now, I’m still better than AI. And more modest too. :)

But now let’s imagine a different situation.

Suppose we need answers in a field where there is not enough evidence to provide a clear answer. In fact, sometimes merely identifying the right questions would already be an enormous breakthrough.

At the same time, there may be so much existing data that processing it properly would require several hundred person-years of work.

I do not want to drift into abstraction, so let’s look at a question closely related to medicine. No warp drives or other unreachable technologies yet. Just one small step beyond where we are now.

We have become reasonably good at managing cardiovascular disease and type 2 diabetes. As a result, in countries where people actually follow medical recommendations, cancer has moved to the top of the list.

In other words, malignant tumours are the next major frontier preventing us from living substantially longer than we do today.

I know perfectly well that the immune system is responsible for allowing malignant cells to escape. It is supposed to destroy them immediately, but at some point it stops doing this with 100% efficiency.

I also understand that it would be extremely useful to determine why certain cells become malignant while others do not, and how this process might be prevented. But that takes us much deeper into the subject. If people are interested, I can return to it in a future post.

So, armed with this basic understanding, I asked AI a simple question:

How could we improve or rejuvenate the immune system? Could we increase overall lifespan using drugs that already exist?

This is not an original idea. Many people have already thought of it, so AI should surely be able to suggest something.

And of course it did.

It suggested rapamycin, together with a few senolytics. Then, as usual, it expanded in every possible direction and added more drugs until the list contained roughly ten compounds.

Fine.

We do not have a huge number of drugs that meaningfully affect immune ageing, nor do we have many plausible life-extension drugs in general. The OpenDrugs database contains only 577 compounds on its shortlist and more than 3,000 on its longlist.

Depending on the definition and counting method, there are approximately 21–34 senolytics. Some sources claim as many as 45.

So AI had not shown me everything. It had merely skimmed the surface.

Naturally, it then produced a long but equally superficial explanation of why it had selected those particular drug combinations rather than others.

And, of course, it explained cellular senescence.

I asked it to go deeper.

Why is senescence so important? What makes this field promising?

I should mention that I have two or three different ways of talking to AI. It supposedly knows who I am and what level of explanation I need, although it constantly forgets.

So, in response to my question, it generated another ten pages of text.

Somewhere near the beginning, it wrote:

“A senescent cell does not divide, but often does not die.”

That sentence irritated me enormously.

I began questioning it the way I would question a student during an oral examination. And I already know that unless I start using profanity, it will not explain the issue properly. It assumes that an ordinary user neither needs nor wants a deeper or more precise answer.

That approach does not work with me.

So, in my usual manner, I started asking increasingly specific questions.

Something like:

“Is cell division really the single defining function of a cell, without which it automatically becomes senescent?”

And then a somewhat more leading question:

“If a cell dies, does it somehow remain senescent? Or is that called normal apoptosis? Because once a cell is dead, it cannot meaningfully be described as senescent anymore, can it?”

People who understand the subject are probably asking themselves why I am doing this at all. Surely it is already obvious that the system is confused and does not properly understand the issue.

The problem is that AI has almost no capacity for self-criticism. It will not correct its reasoning unless you deliberately guide it along the right path—and to do that, you must already know the path yourself.

Ordinary users usually just believe it.

So, for anyone who does not have quite the same intimate relationship with AI that I do, let me explain what happened next.

We went deeper into the subject and eventually identified at least ten direct and indirect mechanisms leading to cellular senescence.

For those interested:

Loss of TRF1, TRF2 or POT1-mediated telomere protection → telomere dysfunction-induced DNA-damage foci → MRN–ATM/ATR–CHEK1/2–TP53–CDKN1A/p21; later, frequently CDKN2A/p16–RB1.
Double-strand DNA breaks: MRE11–RAD50–NBN → ATM → CHEK2. Replication stress: RPA–ATR–CHEK1 → p53–p21 and/or p16–RB.
Hyper-replication → DNA damage and reactive oxygen species; p14ARF/p19Arf–MDM2–p53–p21; p16–RB–E2F.
DNA or telomere damage → ATM–p53–p21; p38 MAPK–p16; NOX2/NOX4 and NF-κB activation; in the presence of advanced glycation end products, also endoplasmic-reticulum stress.
Reduced NAD⁺/NADH availability → AMPK–p53–p21; sometimes reactive oxygen species–induced DNA-damage responses; impaired PINK1/PARKIN and SIRT1/3 signalling.
The unfolded protein response: PERK–EIF2A–ATF4, IRE1–XBP1/JNK and ATF6; reactive oxygen species, p38 and mTORC1 → p21/p16.
Reduced LMNB1, reactivation of repetitive elements and cytoplasmic chromatin fragments → cGAS–STING–TBK1–NF-κB; p16/p21.
IL-1α/β, IL-6, CXCL8 and CCL2 → NF-κB and JAK–STAT signalling; TGF-β–SMAD2/3–p15/p21; contact-dependent NOTCH1–JAG1 signalling.
Usually TGF-β–SMAD–CDKN1A/p21–RB; frequently without p53 activation or a persistent DNA-damage response.
Telomere or genomic damage, p21/p16 activation, p38 signalling, cGAS–STING activation, and mitochondrial and lysosomal dysfunction.

Dasatinib, for example, acts after a cell has already become senescent. It affects several survival mechanisms associated with p21 and/or p16–RB signalling and helps push the senescent cell towards apoptosis, ending its harmful inactivity.

We also discussed the fact that there are no drugs capable of targeting every pathway through which senescent cells arise.

There are certainly no reliably effective interventions of this kind. If they existed, they could potentially become tools for immune rejuvenation—if not immortality—because, in principle, we might be able to create functionally immortal stem-cell populations without simultaneously increasing oncogenic risk.

Then I asked:

What about gene therapy?

What about RNA delivery, CRISPR–Cas9, retroviral vectors, adenoviral vectors and other delivery systems?

Until that moment, AI had not mentioned them at all.

Naturally, the moment I asked, it immediately began explaining how promising all of these technologies were.

This is a good example of a problem in which the amount of available information is simply too large for AI to process properly during an ordinary user session.

That is why we build more complex analytical systems.

I am still not finished with this subject, but the post is becoming far too long.

For now, I would summarise the problem like this:

AI produced a list of mechanisms involved in senescence without understanding which mechanisms were causes, which were consequences, and which were merely associated phenomena. It simply put everything into one large pile.

When you ask it a narrow, highly specific question, it will probably not make a major mistake.

But you have to ask a great many narrow questions.

So, in the next post, I will discuss one possible solution for identifying the most important questions—and, eventually, finding answers to them.

And finally, an old joke that accurately reflects my attitude towards AI’s first answers to ordinary user queries. Reading it is optional.

A group of mice went to a wise owl and asked how they could avoid being eaten by increasingly aggressive cats.

The owl replied:

“Become hedgehogs. If you have spikes, nobody will eat you.”

The mice were overwhelmed by the brilliance of this advice and ran home in excitement.

After a while, however, they came to their senses and returned to the owl.

“Owl, tell us—how exactly do we become hedgehogs?”

The owl replied:

“My job is strategy. I’m not interested in all your fucking tactical details.”

[Yuriy Poteshkin @drpoteshkin]

Let’s continue the discussion about AI and information.

For anyone who did not read the previous post carefully, I’ll repeat the main point: even when AI has access to all the relevant medical literature and all the available patient data, it still cannot consistently make correct clinical decisions based on that information.

That is my opinion, of course. For now, I’m still better than AI. And more modest too. :)

But now let’s imagine a different situation.

Suppose we need answers in a field where there is not enough evidence to provide a clear answer. In fact, sometimes merely identifying the right questions would already be an enormous breakthrough.

At the same time, there may be so much existing data that processing it properly would require several hundred person-years of work.

I do not want to drift into abstraction, so let’s look at a question closely related to medicine. No warp drives or other unreachable technologies yet. Just one small step beyond where we are now.

We have become reasonably good at managing cardiovascular disease and type 2 diabetes. As a result, in countries where people actually follow medical recommendations, cancer has moved to the top of the list.

In other words, malignant tumours are the next major frontier preventing us from living substantially longer than we do today.

I know perfectly well that the immune system is responsible for allowing malignant cells to escape. It is supposed to destroy them immediately, but at some point it stops doing this with 100% efficiency.

I also understand that it would be extremely useful to determine why certain cells become malignant while others do not, and how this process might be prevented. But that takes us much deeper into the subject. If people are interested, I can return to it in a future post.

So, armed with this basic understanding, I asked AI a simple question:

How could we improve or rejuvenate the immune system? Could we increase overall lifespan using drugs that already exist?

This is not an original idea. Many people have already thought of it, so AI should surely be able to suggest something.

And of course it did.

It suggested rapamycin, together with a few senolytics. Then, as usual, it expanded in every possible direction and added more drugs until the list contained roughly ten compounds.

Fine.

We do not have a huge number of drugs that meaningfully affect immune ageing, nor do we have many plausible life-extension drugs in general. The OpenDrugs database contains only 577 compounds on its shortlist and more than 3,000 on its longlist.

Depending on the definition and counting method, there are approximately 21–34 senolytics. Some sources claim as many as 45.

So AI had not shown me everything. It had merely skimmed the surface.

Naturally, it then produced a long but equally superficial explanation of why it had selected those particular drug combinations rather than others.

And, of course, it explained cellular senescence.

I asked it to go deeper.

Why is senescence so important? What makes this field promising?

I should mention that I have two or three different ways of talking to AI. It supposedly knows who I am and what level of explanation I need, although it constantly forgets.

So, in response to my question, it generated another ten pages of text.

Somewhere near the beginning, it wrote:

“A senescent cell does not divide, but often does not die.”

That sentence irritated me enormously.

I began questioning it the way I would question a student during an oral examination. And I already know that unless I start using profanity, it will not explain the issue properly. It assumes that an ordinary user neither needs nor wants a deeper or more precise answer.

That approach does not work with me.

So, in my usual manner, I started asking increasingly specific questions.

Something like:

“Is cell division really the single defining function of a cell, without which it automatically becomes senescent?”

And then a somewhat more leading question:

“If a cell dies, does it somehow remain senescent? Or is that called normal apoptosis? Because once a cell is dead, it cannot meaningfully be described as senescent anymore, can it?”

People who understand the subject are probably asking themselves why I am doing this at all. Surely it is already obvious that the system is confused and does not properly understand the issue.

The problem is that AI has almost no capacity for self-criticism. It will not correct its reasoning unless you deliberately guide it along the right path—and to do that, you must already know the path yourself.

Ordinary users usually just believe it.

So, for anyone who does not have quite the same intimate relationship with AI that I do, let me explain what happened next.

We went deeper into the subject and eventually identified at least ten direct and indirect mechanisms leading to cellular senescence.

For those interested:

Loss of TRF1, TRF2 or POT1-mediated telomere protection → telomere dysfunction-induced DNA-damage foci → MRN–ATM/ATR–CHEK1/2–TP53–CDKN1A/p21; later, frequently CDKN2A/p16–RB1.
Double-strand DNA breaks: MRE11–RAD50–NBN → ATM → CHEK2. Replication stress: RPA–ATR–CHEK1 → p53–p21 and/or p16–RB.
Hyper-replication → DNA damage and reactive oxygen species; p14ARF/p19Arf–MDM2–p53–p21; p16–RB–E2F.
DNA or telomere damage → ATM–p53–p21; p38 MAPK–p16; NOX2/NOX4 and NF-κB activation; in the presence of advanced glycation end products, also endoplasmic-reticulum stress.
Reduced NAD⁺/NADH availability → AMPK–p53–p21; sometimes reactive oxygen species–induced DNA-damage responses; impaired PINK1/PARKIN and SIRT1/3 signalling.
The unfolded protein response: PERK–EIF2A–ATF4, IRE1–XBP1/JNK and ATF6; reactive oxygen species, p38 and mTORC1 → p21/p16.
Reduced LMNB1, reactivation of repetitive elements and cytoplasmic chromatin fragments → cGAS–STING–TBK1–NF-κB; p16/p21.
IL-1α/β, IL-6, CXCL8 and CCL2 → NF-κB and JAK–STAT signalling; TGF-β–SMAD2/3–p15/p21; contact-dependent NOTCH1–JAG1 signalling.
Usually TGF-β–SMAD–CDKN1A/p21–RB; frequently without p53 activation or a persistent DNA-damage response.
Telomere or genomic damage, p21/p16 activation, p38 signalling, cGAS–STING activation, and mitochondrial and lysosomal dysfunction.

Dasatinib, for example, acts after a cell has already become senescent. It affects several survival mechanisms associated with p21 and/or p16–RB signalling and helps push the senescent cell towards apoptosis, ending its harmful inactivity.

We also discussed the fact that there are no drugs capable of targeting every pathway through which senescent cells arise.

There are certainly no reliably effective interventions of this kind. If they existed, they could potentially become tools for immune rejuvenation—if not immortality—because, in principle, we might be able to create functionally immortal stem-cell populations without simultaneously increasing oncogenic risk.

Then I asked:

What about gene therapy?

What about RNA delivery, CRISPR–Cas9, retroviral vectors, adenoviral vectors and other delivery systems?

Until that moment, AI had not mentioned them at all.

Naturally, the moment I asked, it immediately began explaining how promising all of these technologies were.

This is a good example of a problem in which the amount of available information is simply too large for AI to process properly during an ordinary user session.

That is why we build more complex analytical systems.

I am still not finished with this subject, but the post is becoming far too long.

For now, I would summarise the problem like this:

AI produced a list of mechanisms involved in senescence without understanding which mechanisms were causes, which were consequences, and which were merely associated phenomena. It simply put everything into one large pile.

When you ask it a narrow, highly specific question, it will probably not make a major mistake.

But you have to ask a great many narrow questions.

So, in the next post, I will discuss one possible solution for identifying the most important questions—and, eventually, finding answers to them.

And finally, an old joke that accurately reflects my attitude towards AI’s first answers to ordinary user queries. Reading it is optional.

A group of mice went to a wise owl and asked how they could avoid being eaten by increasingly aggressive cats.

The owl replied:

“Become hedgehogs. If you have spikes, nobody will eat you.”

The mice were overwhelmed by the brilliance of this advice and ran home in excitement.

After a while, however, they came to their senses and returned to the owl.

“Owl, tell us—how exactly do we become hedgehogs?”

The owl replied:

“My job is strategy. I’m not interested in all your fucking tactical details.”

[Yuriy Poteshkin @drpoteshkin]

A patient comes to see a doctor and says they have already asked ChatGPT everything.

“If anyone has a problem with that, I personally don’t.”

I like people who make an effort to understand what is happening to them—people who used to search Google and now talk to AI. “My” patients are usually people who value themselves and their health. So if someone has already read about their condition, uploaded their test results, and had a chatbot explain everything, that saves me time and allows us to discuss what actually matters.

“Hold on,” someone in the audience might ask. “If you are not going to explain what the test results mean, what exactly are you going to talk about?”

Historically, doctors have had to begin by explaining what the results show, what each finding means, whether there is a problem, and why the patient should pay attention to it. In my practice, that can take up to 50% of the entire consultation.

The other half is spent discussing what should actually be done—and how.

Modern AI agents have access to an enormous body of knowledge and scientific literature. They also have plenty of manufactured confidence and other stylistic tricks that make people trust them.

But there is a catch.

I have tried to automate patient management and treatment using AI. Trying to replace myself with AI is something of a personal obsession, somehow perfectly compatible with my desire to do as little meaningless work as possible—also known as laziness.

But it did not work.

At least, I was unable to recreate myself.

My clinical reasoning operates across several levels:

Clinical guidelines and algorithms
Large-scale studies
Randomized controlled trials
Individual clinical cases

Traditional literature reviews are often of limited value in clinical practice because they frequently lack a practical direction. They summarize the evidence but do not necessarily help solve a specific clinical problem.

Typical patients can often be managed at level one. But patients whose clinical presentation falls, even partially, outside the scenarios covered by levels one through four find themselves in a frightening situation: the doctor has to make a decision based on the available evidence and their own clinical reasoning.

In other words, the exact decision required has never been described in the literature.

For a doctor, the absence of a clear answer in the literature means that, if the case ever ends up in court, experts will be called in, everything will become complicated, and you may ultimately be judged to have been wrong.

My own rule is to act in the patient’s best interests. Whatever happens afterward is secondary.

I could explore this subject at almost any depth and provide numerous real-life clinical examples, but that deserves a separate journal article.

So what happens when we introduce AI?

My experiments show that even when you provide the system with all the relevant context, including anonymized patient data—and even when you personally find and supply the appropriate guidelines, algorithms, studies, and other evidence—it still struggles to determine what is important and what is secondary, which interventions should come first, and which can wait.

Today’s AI lacks both the breadth and the precision of reasoning required for this task. It also lacks the ability to disregard low-value facts supported by weak evidence.

To an AI system, almost everything written in a credible-sounding source appears true by default. As a result, instead of identifying the few pieces of information that genuinely matter, it sees a large collection of facts and fails to understand their relative significance.

That is why we still do not have a reliable tool capable of fully replacing a physician.

And if someone tells you that they have already built one, check their medical diploma and academic degree. Most likely, at least one of them is missing—possibly both.

Now let us imagine asking a machine a question for which no answer exists in the current medical literature.

We will talk about that in the next post.

[Yuriy Poteshkin @drpoteshkin]

A patient comes to see a doctor and says they have already asked ChatGPT everything.

“If anyone has a problem with that, I personally don’t.”

I like people who make an effort to understand what is happening to them—people who used to search Google and now talk to AI. “My” patients are usually people who value themselves and their health. So if someone has already read about their condition, uploaded their test results, and had a chatbot explain everything, that saves me time and allows us to discuss what actually matters.

“Hold on,” someone in the audience might ask. “If you are not going to explain what the test results mean, what exactly are you going to talk about?”

Historically, doctors have had to begin by explaining what the results show, what each finding means, whether there is a problem, and why the patient should pay attention to it. In my practice, that can take up to 50% of the entire consultation.

The other half is spent discussing what should actually be done—and how.

Modern AI agents have access to an enormous body of knowledge and scientific literature. They also have plenty of manufactured confidence and other stylistic tricks that make people trust them.

But there is a catch.

I have tried to automate patient management and treatment using AI. Trying to replace myself with AI is something of a personal obsession, somehow perfectly compatible with my desire to do as little meaningless work as possible—also known as laziness.

But it did not work.

At least, I was unable to recreate myself.

My clinical reasoning operates across several levels:

Clinical guidelines and algorithms
Large-scale studies
Randomized controlled trials
Individual clinical cases

Traditional literature reviews are often of limited value in clinical practice because they frequently lack a practical direction. They summarize the evidence but do not necessarily help solve a specific clinical problem.

Typical patients can often be managed at level one. But patients whose clinical presentation falls, even partially, outside the scenarios covered by levels one through four find themselves in a frightening situation: the doctor has to make a decision based on the available evidence and their own clinical reasoning.

In other words, the exact decision required has never been described in the literature.

For a doctor, the absence of a clear answer in the literature means that, if the case ever ends up in court, experts will be called in, everything will become complicated, and you may ultimately be judged to have been wrong.

My own rule is to act in the patient’s best interests. Whatever happens afterward is secondary.

I could explore this subject at almost any depth and provide numerous real-life clinical examples, but that deserves a separate journal article.

So what happens when we introduce AI?

My experiments show that even when you provide the system with all the relevant context, including anonymized patient data—and even when you personally find and supply the appropriate guidelines, algorithms, studies, and other evidence—it still struggles to determine what is important and what is secondary, which interventions should come first, and which can wait.

Today’s AI lacks both the breadth and the precision of reasoning required for this task. It also lacks the ability to disregard low-value facts supported by weak evidence.

To an AI system, almost everything written in a credible-sounding source appears true by default. As a result, instead of identifying the few pieces of information that genuinely matter, it sees a large collection of facts and fails to understand their relative significance.

That is why we still do not have a reliable tool capable of fully replacing a physician.

And if someone tells you that they have already built one, check their medical diploma and academic degree. Most likely, at least one of them is missing—possibly both.

Now let us imagine asking a machine a question for which no answer exists in the current medical literature.

We will talk about that in the next post.

[Yuriy Poteshkin @drpoteshkin]

https://t.co/ayD5HpKDVB

[Yuriy Poteshkin @drpoteshkin]

https://t.co/ayD5HpKDVB

[Yuriy Poteshkin @drpoteshkin]

Wow, is it so bad? does not work at all?
Who has the Opus 4.8 to download for the future? :)

[Alexey Strygin @strygah]

New Fable benchmarks from @claudeai

[Yuriy Poteshkin @drpoteshkin]

And if anyone will keep searching for a way through something considered impossible to cure, it is you.

That search may ultimately change how this disease is treated — and hopefully force a serious rethinking of the guidelines.

[Yuriy Poteshkin @drpoteshkin]

When trying to extend your life, never forget that one missed disease can shorten it.

Longevity does not start with supplements, wearables, or protocols.

It starts with excellent clinical medicine.

[Bryan Johnson @bryan_johnson]

Bad news #1:

I have an autoimmune disease. My stomach is eating itself.

Bad news #2:

2–5% of people have this, too. Likely more, because it hides.

Good news:

I'm going to try and solve it. Will share all.

As a kid, I ate sugar cereal, drank sugary soda, and gobbled down fast food. I had a few healthy years in my early 20s but then became a young father of three and began building a business.

Juggling that stress and grind, I let my health slip and gained 40 lbs. Within a few years I’d fallen into a deep, chronic depression.

Somewhere in that timeline, my body began developing an autoimmune process affecting my thyroid and then my stomach lining.

It’s called Autoimmune Gastritis (AIG).

My hypothyroidism got diagnosed when I was 21 years old with a routine blood draw. That enabled me to begin proactive management, supplementing levothyroxine and Armour Thyroid. They are the hormones my body should be producing on its own but wasn’t.

By taking these pills daily, my body was able to operate as though my thyroid was functioning properly. What I didn’t know was that something else was going on inside my body: my stomach had begun attacking itself. But there was no routine test to find out and I didn’t have any symptoms.

I just discovered it in May. I'm unsure how long I've had it. AIG causes irreversible damage: nutritional deficiency, anemia, and over a long horizon, elevated cancer risk. When AIG is discovered today, standard medical care concedes defeat, stating that nothing can be done except managing the condition, no matter how awful or lethal the effects.

Looking back over the past few years, I can now see the early signals we were picking up in measurement but hadn’t connected the dots. For 11 years, I’ve had low ferritin, without anemia. We continually tried to raise my iron levels with food and supplementation but nothing would work.

We chased the obvious solutions first. A plant-based diet means all my iron is the hard-to-absorb, non-heme kind. Hard training, sauna, and hyperbaric oxygen all raise the body's demand for iron. But none of them explained the core failure: despite me taking iron orally, trialing every formulation, and using every timing trick, none of the iron would stick.

What I didn’t fully appreciate until recently is how many stones my previous providers had left unturned. The low ferritin kept getting explained away but not fixed.

I overhauled my medical team earlier this year. It was the rebuild to lay the groundwork for Immortals Care, our $1M a year protocol. With greater capacity, we revisited everything.

On the surface, my low ferritin was easy to dismiss by most standards of care. My hemoglobin and hematocrit were normal. Ferritin measures stored iron, while hemoglobin measures circulating iron, and because the body drains its reserves first to keep hemoglobin normal, you can be fully iron deficient with a perfectly normal hemoglobin and hematocrit.

This is why my low ferritin kept getting dismissed: the numbers that define anemia looked fine, so no one asked why my iron reserves wouldn't refill.

My team pressed on that question. They first turned to a colonoscopy. I was 48 years old and overdue. It was good health hygiene to have while also serving a specific purpose of searching for a hidden source of blood loss such as a polyp or even cancer in my bowels. Either one of those would be an explanation of why the iron kept disappearing.

At the same time, they began connecting the dots. Iron absorption depends on stomach acid, so one theory was that my stomach acid was disrupted. They also knew that thyroid and stomach autoimmunity often travel together, so often that the pairing has a name: thyrogastric syndrome.

Put against my 27+ year history of autoimmune thyroid disease, the pieces pointed to a single hypothesis: my own immune system was attacking my stomach.

To our surprise, my colonoscopy came back clean. A perfectly healthy colon, better than 95% of colonoscopies of men, according to the gastroenterologist. That ruled out the first concern and worst possible outcome: slow continuous bleeding from colon cancer, or pre-cancerous polyp.

My team had exercised great foresight though, anticipating this possible outcome. In addition to a colonoscopy, they’d ordered an upper endoscopy to be performed at the same time. The combined procedure is a bi-directional endoscopy. Probes would look at my entire intestinal tract, up from below and down the throat.

Additionally, we had several blood biomarkers measured ahead of the procedure to try and pick up on any signals that would give the gastroenterologist guidance for what to look for while doing visual inspections.

Fifteen minutes before the procedure, my blood results returned, finding elevated levels of anti-parietal-cells-antibodies (APCA). They came back at roughly five times the upper limit of normal (103, against a ceiling of 20 Units/mL). It was a positive result confirming the suspicion of AIG being the culprit behind my low ferritin, the other type of gastritis, driven by a bacterial infection, was already ruled out, as we knew I am negative to H. pylori.

Even before this finding, my team had ordered five biopsies to be taken from three regions of my stomach.

The biopsies were the critical piece. Had they not been ordered, the bi-directional endoscopy would have been completed and AIG remained undiagnosed as there were no visual signatures of the condition in my intestines.

Two days later, the results of biopsies came in, showing clear signs of early autoimmune gastritis: early atrophy confined to the acid-producing lining, with the rest of the stomach still spared. My team had anticipated this, methodically tracing every line of evidence.

We now had a formal diagnosis. I have autoimmune gastritis AIG. My stomach is eating itself.

So this was never one problem. It was three, linked to one another: the iron deficiency, the autoimmune gastritis driving it, and the autoimmune thyroid disease alongside it. Iron and thyroid feed each other both ways, low iron impairs the conversion of thyroid hormone into its active form, and an under active thyroid impairs how the body uses iron. Each made the other harder to fix.

Autoimmune gastritis affects an estimated 2–5% of people, and likely more, because it hides and is challenging to diagnose. It's usually silent for years, surfacing only once the stomach has atrophied enough to do real damage: iron deficiency first, then B12 deficiency, then anemia from both, and over a long horizon, raised stomach-cancer risk. In one study of people with precancerous gastric lesions, roughly 18% carried the autoimmune antibodies, and only about 1% had ever been diagnosed.

And the earliest clue, low ferritin, is the one standard medicine waves through. Low iron stores get normalized and rarely investigated at all when anemia hasn't shown up yet. That blind spot is what hid mine for a decade.

The good news: the iron deficiency is now corrected. I received a 1,000 mg Monoferric iron infusion. This was chosen for two reasons after considering multiple formulations. First, it can safely deliver a full dose of iron in a single infusion (1,000 mg), while older options like Venofer require several separate appointments to reach the same total.

Second, certain other IV iron formulations can cause a drop in blood phosphate levels, an important mineral for bones and energy. Monoferric is much less likely to do this, which matters given how closely we track long-term metabolic and bone health parameters.

As mentioned earlier, current medical standards treat AIG as something to be managed, not resolved.

It's worth noting that many of you give me a hard time, inviting me to "live life" and engage in self-destructive behaviors like a "normal person". I'm cool with the playful ribbing. Also, had I not taken care of my health during the past five years, my situation could potentially be very serious.

You too may have a lurking health issue that is undiagnosed and could increase in severity from unhealthy life choices, without your knowing. The absence of symptoms is not the presence of health.

A gentle nudge that minding your health, no matter your situation in life, is good decision making.

My team and I are going to try and solve my AIG. This is how we’re approaching it:

First, routine monitoring keeps the disease in view: ferritin and iron, B12, the pepsinogen I/II ratio, gastrin, and chromogranin A. Gastrin is the dial to watch. If it climbs, the disease is advancing, and the risk of gastric neuroendocrine tumors climbs with it.

Second, we’re doing advanced characterization of the disease. We’ll do a repeat biopsy to read the immune infiltrate, deep cytokine profiling, and T-cell subset analysis, to see which pathways are actually firing.

That testing drives the intervention plan, including the experimental approaches we intend to develop.

+ If gastrin and chromogranin rise: damp the gastrin drive (netazepide) and tighten endoscopic surveillance. If the profile is Th1 / interferon-driven: target JAK/STAT.

+ If it's Th17 / IL-17-driven: target IL-17 and STAT3.

+ If regulatory T cells are failing: rebuild them (low-dose IL-2, induced Tregs).

+ If it's antibody- and B-cell-driven and antigen-specific: engineered cell therapy (CAAR-T).

Which organizes into four tiers, from available today to frontier:

Tier 1, now: protect and support; zinc-L-carnosine, and acid replacement (betaine HCl with pepsin) under physician supervision. This is specific to my case and not something to self-prescribe, especially given the cancer-surveillance considerations above.

Tier 2, target the signaling , JAK/STAT, GSK-3, IL-17, and damp the gastrin drive (netazepide).

Tier 3, reset the cells, induced regulatory T cells (iTregs).

Tier 4, frontier: engineered T-cell therapy (CAR-T / CAAR-T), custom AI-designed antibodies, or synthetic proteins, that can specifically seek out inactivate or destroy the rogue immune cells attacking my stomach lining.

To be clear: there's no approved cure for autoimmune gastritis today. Medicine treats it as something to manage, not solve. Tiers 2 through 4 are investigational preclinical evidence at best, and in several cases therapies that still have to be built.

If you're working on autoimmune gastritis, antigen-specific tolerance, regulatory T cells, or CAAR-T for organ-specific autoimmunity, please reach out.

Modern medicine has normalized too many conditions that erode our health, function, and comfort, shrinking the goal to monitoring and management while a cure is rarely even attempted. Most of these verdicts were handed down decades ago, in an era that predates nearly all of our current tech and science, and they have gone largely unchallenged.

We want to change that. In the age of AI, multiomics, and custom-built DNA, proteins, and cells, no condition should be presumed incurable simply because no one has yet tried to cure it with today's stack.

I’ll end on a personal note.

We fill our days mostly on things that are trivial next to what we ultimately care about. We know, deep down, however, that in the noise of it all, health is easily forgotten until it’s the only thing that matters.

We spend a fraction of our lives truly sober to the preciousness of life. We feel it when someone we love dies, when a child is born, when we come close to death ourselves, or when a diagnosis marks our limit. In those moments, we are sobered, and the rarity of it all becomes self evident. Imagine the existence we’d build together if that clarity didn’t fade.

I wish all of you the very best. Care for yourself, care for others, care for the planet and care for our animal friends. Care for life as it’s the most precious gift there is.

[Yuriy Poteshkin @drpoteshkin]

That combination is rare, because many great clinicians are still skeptical of the longevity movement.

@bryan_johnson Bryan, I hope you beat this disease.

Your choice of CAR-T as a mechanism-based treatment strategy looks logical and comparatively accessible today.

[Yuriy Poteshkin @drpoteshkin]

Grok is the only big LLM company that allows to work with the censored by other players data, this is a huge advantage to me, as medicine is one of the censored fields, hope imagine will be getting better to be able to visualize more complicated medical stuff.

[Elon Musk @elonmusk]

Made with Grok Imagine

[Yuriy Poteshkin @drpoteshkin]

We really should fight death. In the past people thought it is ok to die from the pneumonia and they did not understand the real cause of the disease. Right now we do not fully understand why people die anyway. We should know more and fund the science for this more.

[Julian Issa @juliankissa]

Our vitalist / longevity special with @realNathanCheng and @adamgries is out now on The Beyond Tomorrow Podcast

[Yuriy Poteshkin @drpoteshkin]

and how many additional years of life we could gain by choosing the right path. That is the direction I want Eternal Search to move in: from mapping separate ideas to understanding the route toward real life extension. #eternalsearch

[Yuriy Poteshkin @drpoteshkin]

Another update on https://t.co/TTvEz5H3t1

Working with ideas has turned out to be surprisingly interesting.

There are many of them. Some repeat under different angles and need deduplication. That is the easy part.

The more interesting part is how ideas relate to each other.

[Yuriy Poteshkin @drpoteshkin]

Some contradict each other. Some open new possibilities. Some need to happen together before a third direction becomes possible. Over time, this flow should become a clearer picture of the future: what has to happen, in what sequence, how we get there, how much it may cost,

[Mikhail Batin @MikhailBatin]

I'm proud to introduce our new project: Eternal Search.

Eternal Search is a deep understanding of the field and a search for the best solutions for the survival of humans and humanity. Eternal Search: we exist so that you exist.
It matters to all of us to understand how the fight against death actually works. We want to live in a world where saving lives comes first.

Right now our database holds 1,037 organizations, 666 grants, 4,425 investors (1,019 of them with complete dossiers), 24 longevity companies (at least 140 signals tracked on each), 70 longevity projects, 41 key figures, around 4,000 public pages and 12,000 internal ones.

At this moment our AI agents are processing and scoring 2,000 ideas about extending life — ideas put forward by companies, by leading figures, and by academic science.
In two months, we've published 772 news items from the field.

Grants, investors, funds, labs, research groups, projects, key people, and companies — we see who's working on what, who's worth watching, and where it's all heading. And we mean to grow.

We see the battle of ideas in life extension, and we're in it. We're looking for the answer to the question: where should the money, the talent, and the attention go to raise our odds of staying alive?

We see every active grant — and for each one, we know who it's truly right for. Come find the money with us.

Using both natural and artificial intelligence, we're analyzing 2,750 of the core questions of transhumanism right now. We can't afford to get it wrong by choosing the wrong question to answer.

Join us — as a developer, a volunteer, or a patron — and help build the best longevity analytics in the world. We have to break through the Longevity Winter, beat the records for lifespan, and find a way for people to survive.

Make your decisions on good data and a deep understanding of the situation

https://t.co/lVBacM9vZF

[Yuriy Poteshkin @drpoteshkin]

Biology is currently dominated by regulation. Technology/methodology is dominated by discovery.
Next: connect ideas with companies, projects, founders, scientists, and intellectual influences — so we can see who is building on what, which directions are strong
#eternalsearch

[Yuriy Poteshkin @drpoteshkin]

https://t.co/fpPG35jTna is 70 days old today. 70 straight days with commits.

Right now we are building an idea navigator for longevity.

Current map:
15 core ideas
179 sub-ideas
The taxonomy is bottom-up: real companies and projects first, classification after.